Researchers propose 'computational identifiability,' a new framework that redefines how causal effects are identified in data science by shifting from theoretical, infinite-data assumptions to practical, finite computational search procedures. This approach enables identification under realistic conditions including small samples, ambiguous graphical criteria, and mixed observational-interventional data.
The distinction between theoretical and computational identifiability addresses a longstanding gap in causal inference research. Traditional identifiability assumes asymptotic properties and infinite data, creating a disconnect between what mathematics proves possible and what practitioners can actually achieve with finite computational resources. This paper bridges that gap by proposing a framework where identifiability is satisfied if a computational search procedure successfully finds an estimator within specified error tolerances, conditional on defined assumptions and search parameters.
The computational identifiability framework emerges from growing recognition that real-world data science operates under severe constraints absent from classical theory. Researchers and practitioners frequently encounter small finite samples, ambiguous causal structures, and mixed data types that violate textbook assumptions. Previous approaches either provided binary yes/no identifiability answers disconnected from practical implementation or offered no guidance for these nuanced scenarios.
For the broader AI and data science community, this work has significant implications. Machine learning practitioners working with causal models can now ask fine-grained questions about identifiability tailored to their specific computational constraints and data characteristics. The framework enables more honest assessment of what can be reliably estimated in real scenarios, potentially preventing overconfident causal claims built on theoretical assumptions that don't hold empirically.
The open-source code release democratizes access to these methods, allowing practitioners to test computational identifiability across their own problems. Future applications may include improved causal inference in limited-data domains like medical research, economics, and policy evaluation where both theoretical guarantees and practical computational feasibility matter.
- →Computational identifiability shifts from infinite-data theoretical assumptions to practical finite-sample search procedures for finding causal estimators.
- →The framework enables identification assessment under realistic conditions including small samples, ambiguous causal structures, and mixed observational-interventional data.
- →Identifiability becomes conditional on specified search assumptions and prior distributions, making it context-dependent rather than universally true.
- →Open-source implementation allows practitioners to evaluate identifiability across diverse real-world scenarios beyond classical theory.
- →This approach bridges the gap between what mathematics proves theoretically possible and what computational resources actually achieve in practice.